Evaluating multimodal AI in medical diagnostics

npj Digital Medicine, Journal Year: 2024, Volume and Issue: 7(1)

Published: Aug. 7, 2024

This study evaluates multimodal AI models' accuracy and responsiveness in answering NEJM Image Challenge questions, juxtaposed with human collective intelligence, underscoring AI's potential current limitations clinical diagnostics. Anthropic's Claude 3 family demonstrated the highest among evaluated models, surpassing average accuracy, while decision-making outperformed all models. GPT-4 Vision Preview exhibited selectivity, responding more to easier questions smaller images longer questions.

Language: Английский

---

AI-assisted radiologists vs. standard double reading for rib fracture detection on CT images: A real-world clinical study

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(1), P. e0316732 - e0316732

Published: Jan. 24, 2025

To evaluate the diagnostic accuracy of artificial intelligence (AI) assisted radiologists and standard double-reading in real-world clinical settings for rib fractures (RFs) detection on CT images. This study included 243 consecutive chest trauma patients (mean age, 58.1 years; female, 166) with scans. All scans were interpreted by two radiologists. The images re-evaluated primary readers AI assistance a blinded manner. Reference standards established musculoskeletal re-evaluation results then compared those from initial double-reading. analysis focused demonstrate superiority AI-assisted sensitivity noninferiority specificity at patient level, to Secondary endpoints lesion levels. Stand-alone performance was also assessed. influence characteristics, report time, RF features investigated. At significantly improved 25.0% (95% CI: 10.5, 39.5; P < 0.001 superiority), double-reading, 69.2% 94.2%. And, diagnosis (100%) noninferior (98.2%) difference 1.8% -3.8, 7.4; = 0.999 noninferiority). both influenced gender, number, fracture location, type. Radiologist affected whereas AI’s age side involved. additional-reader workflow might be feasible strategy instead traditional potentially offering higher real-word practice.

Language: Английский

---

Deep Learning Techniques for the Dermoscopic Differential Diagnosis of Benign/Malignant Melanocytic Skin Lesions: From the Past to the Present

Bioengineering, Journal Year: 2024, Volume and Issue: 11(8), P. 758 - 758

Published: July 26, 2024

There has been growing scientific interest in the research field of deep learning techniques applied to skin cancer diagnosis last decade. Though encouraging data have globally reported, several discrepancies observed terms study methodology, result presentations and validation clinical settings. The present review aimed screen literature on application DL dermoscopic melanoma/nevi differential extrapolate those original studies adequately by reporting a model, comparing them among clinicians and/or another architecture. second aim was examine together according standard set statistical measures, third provide dermatologists with comprehensive explanation definition most used artificial intelligence (AI) better/further understand this topic and, parallel, be updated newest applications medical dermatologic field, along historical perspective. After screening nearly 2000 records, subset 54 selected. Comparing 20 convolutional neural network (CNN)/deep (DCNN) models, we scenario highly performant algorithms, especially low false positive results, average values accuracy (83.99%), sensitivity (77.74%), specificity (80.61%). Looking at comparison diagnoses (13 studies), main difference relies values, +15.63% increase for CNN/DCNN models (average 84.87%) compared humans 64.24%) 14,85% gap accuracy; were comparable (79.77% 79.78% humans). To obtain higher diagnostic feasibility practice, rather than experimental retrospective settings, future should based large dataset integrating images relevant anamnestic that is prospectively tested physicians.

Language: Английский

---

Distribution, Recognition, and Just Medical AI

Philosophy & Technology, Journal Year: 2025, Volume and Issue: 38(1)

Published: Jan. 27, 2025

Language: Английский

---

Levelling up as a fair solution in AI enabled cancer screening

Frontiers in Digital Health, Journal Year: 2025, Volume and Issue: 7

Published: Feb. 25, 2025

With the proliferation of AI enabled tools in healthcare, clinicians have raised concerns about potential for bias and subsequent negative impacts on underrepresented groups (1). The causes model deployment are multifaceted can occur throughout development process (2). A well-recognized example within training includes unrepresentative datasets that limit generalizability real-world populations (3). Whilst may outperform current standards care well-represented groups, models perform worse under-represented (4)(5)(6)(7)(8)(9). Diversifying is obvious solution to caused by homogenous data, however data collection a long term project take years or even decades acquire (10). dilemma policymakers currently releasing unfair harm whilst withholding them would cause significant welfare opportunity costs groups. To address this problem, bioethicists like Vandersluis Savalescu (11) suggested alternate strategies such as 'selective deployment', which deploy only This incurs an fairness cost. issue diagnostic testing not specific applications debates also remain ongoing field mainstream medicine how challenge (12,13). By examining case study FIT testing, has been shown more effectively male patients bowel cancer screening, paper supports use sex adjustment 'level up' female (14). Through medicine, lessons learned from policy be transferred clinical will illustrated using parallel AI-assisted breast screening.Bowel screening:NHS England introduced screening detection 2019, see Figure 1 summary workflow (15). tests stool samples measure concentration blood stool, if above specified threshold triggers referral further investigation usually colonoscopy. In UK, National Screening Committee (NSC) set (120µg/g) based cost-effectiveness analysis, where effectiveness determined Quality-Adjusted Life Years (QALY) gained, system capacity (16). lower results 'positive' tests, with greater rate unnecessary colonoscopies. Conversely, higher thresholds result burden colonoscopy constraints at risk missed cancers. There increasing evidence FITs patients, who median faecal measurement than males (17). For each threshold, rates found subgroup (18). Despite this, UK continues universal contrast some other countries adopted sex-adjusted resulting positive test (19). example, Sweden's positivity 40µg/g 80µg/g females respectively, resulted equal proportions subgroups but cost colonoscopies (20). Similar trends seen Finland (21).A cost-benefit analysis conducted. aim assess 'levelling-up' through ensuring equitable health outcomes when 'unfair' being utilised. Following way levelling up applied deployments, assisted mammogram interpretation, explored. These studies illustrate usefulness transfer learning emerging algorithmic fairness.In fairest strategy maximal utility. From public perspective, increased levels similar men reduce overall mortality morbidity (22). Economically, reducing false earlier effective presentations amenable treatment particularly important publicly funded (23). turn, service reduced first line treatments, social associated advanced cancer. Health gains vary between due differences underlying population risk, Sweden shows nearly 25% classified thresholds, were subsequently diagnosed ethical standpoint, acknowledges processes suboptimal fair outcomes, much base grounded white normativity (24).The disadvantaged centre positives. Firstly, posed positives differ depending application, completely free they do constitute relatively intervention. randomised trial exploring effect approximately 12,000 had colonoscopy, there no perforations deaths 30 days post procedure (25). Furthermore, specialist nurses screen all before ensure fit enough safety net mitigate (26). Secondly, healthcare providers able provide necessary unit space, equipment qualified personnel (27). Some argue it increase keeping same could lead equivalent performance without exceeding existing capacity. Although widely accepted down-levelling group unethical. Instead, must mandate prior responsible design approach, so appropriate both considered.Levelling presents unique challenges medicine. advocates adjustments poor subgroups, need post-deployment evaluation complex nature disease manifestation. Data drift refers changes properties over time what was used (28). AI, phenomena whereby mean longer appropriate; presenting younger declining incidence (29). Therefore, should eternally fixed regular conducted continuing purpose meeting ever-changing needs patients.Whilst disadvantages wellrepresented non-white racial benefit contexts. Race-adjustment difficult adopt sexadjustment controversies surrounding race-based stemming historically exploitative practices Sims' experimentation enslaved black women (30). race construct critical historic underpinnings categories continue defined (31). However, race-adjustment useful tool addressing inequality instances uplevel receive inadequate failures, part systemic racism, rather propagate belief innate biological differences. When poorly performing adjustment, transparency understanding trust providers, faced injustice.Though differential one-size-fits-all solution. Rather, intended add research deploying models, allowing comprehensive guide draw from. conditions under most suitable approach. mitigation identifies bias, specifically underdiagnosis. teams understand sequalae referral. different contexts repercussions, application deployed across separate NHS trusts varying guidelines. Adjustment preferred low-risk intervention, high gain screening. Next, workflows human-in-the-loop over-referral, nurse contact fitness testing.An these apply Similarly mammograms offered national programme NHS, workflow. imaging, known computer vision, popular (32). NSC finding lack introduce already begun trialing second reader prospective (33,34). Recent highlighted commercially available diagnosing suspicious lesions images overpredicts (35). raised, context strategy. An intentional initial underdiagnosis subgroup. Levelling given possible low double read requirement (36). acts one readers clinician-in-the-loop query diagnoses seek third opinion necessary. fails (i.e. human wrongly classify suspicious), urgent specialty review organised decide biopsy lowering patients.Levelling doesn't solve reasons why differentially, does offer happen point pipeline. As such, cross functional including developers, researchers attempt elicit act together highlight interventions counteract preventable root causes. initiatives engage efforts techniques synthetic diversify (37)(38)(39).In summary, approach safely balances utility certain met. highlights solutions diagnostics base, AI. harm, essential remains focus bias.

Language: Английский

---

The multifaceted role of phosphodiesterase 4 in tumor: from tumorigenesis to immunotherapy

Frontiers in Immunology, Journal Year: 2025, Volume and Issue: 16

Published: March 10, 2025

Phosphodiesterase 4 (PDE4) is an enzyme that specifically hydrolyzes the second messenger cAMP and has a critical role in regulation of variety cellular functions. In recent years, PDE4 attracted great interest cancer research, its tumorigenesis development been gradually elucidated. Research indicates abnormal expression or heightened activity associated with initiation progression multiple cancers, including lung, colorectal, hematological by facilitating cell proliferation, migration, invasion, anti-apoptosis. Moreover, also influences tumor immune microenvironment, significantly evasion suppressing anti-tumor responses, reducing T-cell activation, promoting polarization tumor-associated macrophages toward pro-tumorigenic phenotype. However, family may have both oncogenic tumor-suppressive effects, which could depend on specific type grade tumor. inhibitors garnered substantial as potential anti-cancer therapeutics, directly inhibiting growth restoring surveillance capabilities to enhance clearance cells. Several are currently under investigation aim exploring their therapy, particularly combination strategies checkpoint inhibitors, improve therapeutic efficacy mitigate side effects conventional chemotherapy. This review provides overview tumorigenesis, drug resistance, immunotherapy, actions intending guide exploration new target therapy.

Language: Английский

---

Artificial Intelligence Performance in Image-Based Cancer Identification: Umbrella Review of Systematic Reviews

Journal of Medical Internet Research, Journal Year: 2025, Volume and Issue: 27, P. e53567 - e53567

Published: April 1, 2025

Background Artificial intelligence (AI) has the potential to transform cancer diagnosis, ultimately leading better patient outcomes. Objective We performed an umbrella review summarize and critically evaluate evidence for AI-based imaging diagnosis of cancers. Methods PubMed, Embase, Web Science, Cochrane, IEEE databases were searched relevant systematic reviews from inception June 19, 2024. Two independent investigators abstracted data assessed quality evidence, using Joanna Briggs Institute (JBI) Critical Appraisal Checklist Systematic Reviews Research Syntheses. further in each meta-analysis by applying Grading Recommendations, Assessment, Development, Evaluation (GRADE) criteria. Diagnostic performance synthesized narratively. Results In a comprehensive analysis 158 included studies evaluating AI algorithms noninvasive across 8 major human system cancers, accuracy classifiers central nervous cancers varied widely (ranging 48% 100%). Similarities observed diagnostic head neck, respiratory system, digestive urinary female-related systems, skin, other sites. Most meta-analyses demonstrated positive summary performance. For instance, 9 meta-analyzed sensitivity specificity esophageal cancer, showing ranges 90%-95% 80%-93.8%, respectively. case breast detection, calculated pooled within 75.4%-92% 83%-90.6%, Four reported ovarian both 75%-94%. Notably, lung was relatively low, primarily distributed between 65% 80%. Furthermore, 80.4% (127/158) high according JBI Checklist, with remaining classified as medium quality. The GRADE assessment indicated that overall moderate low. Conclusions Although shows great achieving accelerated, accurate, more objective diagnoses multiple there are still hurdles overcome before its implementation clinical settings. present findings highlight concerted effort research community, clinicians, policymakers is required existing translate this into improved outcomes health care delivery. Trial Registration PROSPERO CRD42022364278; https://www.crd.york.ac.uk/PROSPERO/view/CRD42022364278

Language: Английский

---

The doctor and patient of tomorrow: exploring the intersection of artificial intelligence, preventive medicine, and ethical challenges in future healthcare

Frontiers in Digital Health, Journal Year: 2025, Volume and Issue: 7

Published: April 3, 2025

Artificial intelligence (AI) 's rapid integration into healthcare transforms medical decision-making, preventive strategies, and patient engagement. AI-driven technologies, including real-time health monitoring predictive analytics, offer new personalized care possibilities. However, concerns regarding ethical implications, data security, equitable access remain unresolved. This paper addresses the critical gap in AI healthcare, highlighting statistical evidence of its impact. It also explores intersection AI, medicine, challenges future envisioning evolving roles physicians patients an AI-integrated ecosystem. A fictional case study projected for 2040, illustrating entirely digitized, AI-supported system, frames discussion about digital privacy regulations, AI's implications medicine. Digital interventions powered by will facilitate strengthen autonomy, enhance precision algorithmic bias, privacy, equity must be addressed to ensure fosters inclusivity rather than exacerbating disparities. Regulatory frameworks, such as GDPR, provide foundational protections, but further adaptations are required govern expanding role digital-assisted medicine has potential redefine patient-provider interactions, efficiency, promote proactive management. achieving this vision requires a multidisciplinary approach involving professionals, policymakers, technology developers. Future research should focus on regulatory literacy, implementation balance innovation with equity, ensuring that remains patient-centered inclusive.

Language: Английский

---

The Use of Artificial Intelligence for Skin Cancer Detection in Asia—A Systematic Review

Diagnostics, Journal Year: 2025, Volume and Issue: 15(7), P. 939 - 939

Published: April 7, 2025

Background: Artificial intelligence (AI) developed for skin cancer recognition has been shown to have comparable or superior performance dermatologists. However, it is uncertain if current AI models trained predominantly with lighter Fitzpatrick types can be effectively adapted Asian populations. Objectives: A systematic review was performed summarize the existing use of artificial detection in Methods: Systematic search conducted on PubMed and EMBASE articles published regarding amongst Information study characteristics, model outcomes collected. Conclusions: Current studies show optimistic results utilizing Asia. comparison image abilities might not a true representation diagnostic versus dermatologists real-world setting. To ensure appropriate implementation, maximize potential AI, improve transferability across various genotypes cancers, crucial focus prospective, real-world-based practice, as well expansion diversification databases used training validation.

Language: Английский

---

Artificial Intelligence to predict cancer risk, are we there yet? a comprehensive review across cancer types

European Journal of Cancer, Journal Year: 2025, Volume and Issue: unknown, P. 115440 - 115440

Published: April 1, 2025

Language: Английский

---